Accelerating Electromagnetic Inverse-Design using Deep Learning

One of the primary ways that deep learning has been applied to electro magnetics in recent years is for accelerating inverse design. We present one such method for designing metasurface supercells which are robust to structural erosion and dilation, a typical variety of nanofabrication error. A pair of deep neural networks are trained to high accuracy to predict diffraction efficiencies from a supercell mask, and then evaluated exhaustively to find tolerance bounds for freeform supercell designs.